vocabulary: runloop-ai
version: 0.1.0
provider: Runloop
description: >
  Operational and capability vocabulary for the Runloop platform. Runloop offers Devboxes (isolated
  Linux sandboxes), Blueprints (image templates), Snapshots, Agents, Axons (event streams),
  Benchmarks/Scenarios (eval framework), Objects (storage), Secrets, Network Policies, Gateway
  Configs, and MCP Configs as the building blocks for AI coding agents.
created: 2026-05-25
modified: 2026-05-25

concepts:
  - id: devbox
    label: Devbox
    description: An isolated Linux microVM-backed sandbox environment with shell, filesystem, networking, and optional GPU. The primary execution unit for AI coding agents.
    aliases: [sandbox, dev box, code sandbox]
  - id: blueprint
    label: Blueprint
    description: Dockerfile-based container image template used to launch Devboxes with pre-installed dependencies and tools.
    aliases: [image template, dev image]
  - id: snapshot
    label: Snapshot
    description: Saved disk state from a Devbox that can be restored into new Devboxes to branch or resume agent work.
    aliases: [disk snapshot, devbox snapshot]
  - id: agent
    label: Agent
    description: Registered, versioned agent definition (sourced from Git, npm, pip, or storage objects) that can be mounted to Devboxes for reproducible installs.
  - id: axon
    label: Axon
    description: Distributed event stream that sequences, records, and observes agent interactions and broker turn-taking.
  - id: broker
    label: Broker
    description: Bridge that connects Axon event streams to agents running inside Devboxes, handling turn-based interaction protocols (e.g. ACP).
  - id: benchmark
    label: Benchmark
    description: Eval framework definition composed of scenarios, scorers, and a scoring contract used to measure agent performance.
  - id: scenario
    label: Scenario
    description: Atomic eval unit consisting of input context, expected outcomes, and a scorer.
  - id: tunnel
    label: Tunnel
    description: HTTPS URL that exposes a Devbox port to the public internet for live previews and webhook reception.
  - id: storage-object
    label: Storage Object
    description: Managed file with a presigned download URL that can be mounted into Devboxes.
  - id: secret
    label: Secret
    description: Encrypted credential stored at account scope and injected at Devbox runtime without exposing the raw value to agent code.
  - id: gateway-config
    label: Gateway Config
    description: Agent Gateway proxy configuration that lets Devbox-resident agents call external AI APIs without seeing the underlying provider credentials.
  - id: mcp-config
    label: MCP Config
    description: Declarative Model Context Protocol server binding that Runloop mounts to Devboxes so agents can call MCP tools.
  - id: network-policy
    label: Network Policy
    description: Egress allow/deny ruleset applied to a Devbox to restrict outbound network access.
  - id: pty-session
    label: PTY Session
    description: Interactive pseudo-terminal session attached to a Devbox via WebSocket data plane; supports reconnect by stable session name.
  - id: execution
    label: Execution
    description: One invocation of a command on a Devbox; can be synchronous (`exec`) or asynchronous (`exec_async`) with streamable stdout/stderr.

services:
  - id: rest-api
    label: REST API
    base_url: https://api.runloop.ai
    description: Primary HTTP/JSON REST API surface for all Runloop resources.
  - id: python-sdk
    label: Python SDK
    package: runloop-api-client
    description: Async-first Python SDK generated by Stainless from the OpenAPI spec.
  - id: typescript-sdk
    label: TypeScript SDK
    package: '@runloop/api-client'
    description: Stainless-generated TypeScript SDK for Node and the browser.
  - id: rl-cli
    label: rl-cli
    description: Interactive Runloop CLI with rich TUI and scripting modes.
  - id: mcp-server
    label: Runloop MCP Server
    description: Model Context Protocol server that exposes Runloop operations as MCP tools to Claude, Cursor, and other MCP-aware hosts.
  - id: dashboard
    label: Runloop Dashboard
    description: Web UI for managing Devboxes, Blueprints, Snapshots, and benchmarks.
  - id: github-actions
    label: GitHub Actions Integration
    description: deploy-agent and supporting Actions for CI/CD-driven agent deployment and benchmark runs.

tools:
  - SWE-Bench Verified runner
  - SWE-smith runner
  - Custom benchmark runner
  - Reinforcement Fine-Tuning (RFT) experiments
  - Supervised Fine-Tuning (SFT) experiments
  - Snapshot branching
  - Suspend/resume

standards:
  - id: openapi-3.0.3
    label: OpenAPI 3.0.3
  - id: mcp
    label: Model Context Protocol
    description: Open Anthropic-led protocol for AI tool integration.
  - id: acp
    label: Agent Communication Protocol
    description: Protocol used by Runloop brokers to bridge Axon event streams to agents.
  - id: soc2-type-2
    label: SOC 2 Type II
  - id: hipaa
    label: HIPAA
  - id: gdpr
    label: GDPR

tags:
  - AI
  - AI Agents
  - Sandboxes
  - Devboxes
  - Coding Agents
  - Evaluation
  - Benchmarks
  - SWE-Bench
  - MCP
  - Snapshots
  - microVM